Modeling Relational Time Series using Gaussian Embeddings


We address the problem of modeling multiple simultaneous time series where the observations are correlated not only inside each series, but among the different series. This problem happens in many domains such as ecology, meteorology, etc. We propose a new dynamical state space model, based on representation learning, for modeling the evolution of such series. The joint relational and temporal dynamics of the series are modeled as Gaussian distributions in a latent space. A decoder maps the latent representations to the observations. The two components (dynamic model and decoder) are jointly trained. Using stochastic representations allows us to model the uncertainty inherent to observations and to predict unobserved values together with a confidence in the prediction.

NIPS Time Series Workshop